Shownotes
A long look at what happens when you apply LLM interpretability tooling to a weather model: sparse autoencoders surface human-legible features inside GraphCast, and nudging the tropical-cyclone feature by hand steers a simulated hurricane in physically consistent ways.

